Viral, Quality, and Junk Videos on YouTube: Separating Content From Noise in an Information-Rich Environment R. Crane and D. Sornette Chair of Entrepreneurial Risks Department of Management, Technology and Economics ¨ ¨ ETH-Zurich, CH-8032 Zurich, Switzerland
Introduction
sense by the collective actions of the community (by mak-
With the rise of web 2.0 there is an ever-expanding source
ing a video the `most-viewed'). We nd that videos chosen
of interesting media because of the proliferation of usergenerated content. However, mixed in with this is a large amount of noise that creates a proverbial needle in the haystack when searching for relevant content.
Although
there is hope that the rich network of interwoven metadata
by the editors (exogenous) have a strikingly different history than those chosen by the community (endogenous). While both classes show a power-law relaxation (inset) over onehundred days following the peak, the videos featured by the community clearly display signi cant precursory growth.
may contain enough structure to eventually help sift through ular things.
Most Viewed Today Front Page
Identifying only the most popular items can be useful, but doing so fails to take into account the famous long tail small, niche interests can outweigh the market share of the few blockbuster (i.e. most-popular) items thus providing only content that has mass appeal and masking the interests of the idiosyncratic many. YouTube, for example, hosts over 40 million videos enough content to keep one occupied for more than 200 years.
Are there intelligent tools to search through this
information-rich environment and identify interesting and
8
10 Aggregate Daily View Count
behavior of the web the notion that the collective effect of
Aggregate Daily View Count
this noise, currently many sites serve up only the most popMost Viewed Today Front Page 8
10
7
10
6
10
2
1
4
6 8
2
4
6 8
10 100 Time (days following peak)
7
10
6
10
relevant content? Is there a way to identify emerging trends or hot topics in addition to indexing the long tail for content that has real value?
0
Information about quality is contained in the dynamics
50
100 150 200 250 300 Time (days centered on peak)
350
Figure 1: A non-parametric superposition of all videos appearing on the `front page' (editorial featuring) and the
We demonstrate that this is possible based on a form of dy-
`most-viewed today' page (community featuring). One im-
namic ltering. In essence, the relaxation signature follow-
mediately sees the exogenous effect of editorial featuring,
ing a burst of viewing activity reveals information about the
revealed by the lack of precursory growth in the view count,
quality of the content. This signature depends on the sus-
whereas endogenous growth is seen in the case of commu-
ceptibility of the social network, in addition to the type of
nity featuring. Inset: power-law relaxation in the 100 days
perturbation that generated the burst.
following the peak reveals long-memory effects.
We begin by considering two classes of perturbations: endogenous and exogenous. Their distinction which is not required to be known a priori is illustrated in gure 1. Here we show the aggregate time-series for videos appearing on the front page of YouTube along with those appearing on the `most-viewed today' page. Videos appearing on the front-page are chosen by the editors, whereas those on the `most-viewed today' page are `chosen' in a collaborative
Once a burst of activity has been triggered, its relaxation depends on the susceptibility of the underlying social network. If the community is ripe for the content, then each generation of viewers can easily pass on the video to the next generation, and one will nd the view count relaxes slowly. If instead the community is uninterested , then even a well-
c 2008, Association for the Advancement of Arti cial Copyright °
orchestrated marketing campaign will fail to spread through
Intelligence (www.aaai.org). All rights reserved.
the network and one will witness a fast relaxation.
Description of Data and Model
6
These ideas have been formalized and tested using a mas-
3
5
Daily View Count
Daily View Count
sive database tracking the time-series of the daily views over 1 year for almost 5 million videos on the popular site
2
1
10
1000
YouTube.com.
10
4
10
5 8 6
an epidemic branching process on a social network.
count of a video results from many factors such as featuring on YouTube, emailing (or other forms of sharing videos), embedding and linking from external websites, discussion
10
30
40
4 2 4
4
10 0
10
20 30 Time (days)
40
50
This
book sales (Sornette et al. 2004). The instantaneous view
1
10
5
dynamical response of the daily view count in the context of model was previously applied successfully to the case of
5
2
8 7 6
Quantifying these effects can be achieved by studying the
6
6 8 6
10
4
Figure 2:
0
10
20
50
Time (days)
Examples of endogenous (left) and exogenous
(right) bursts of activity for individual videos. The exponent of the power-law relaxation (inset) can be used to classify videos as viral, quality, or junk.
on blogs, in newspapers, television, and from social effects in which viewers may be in uenced by others in their network. The impact of these various factors may not be immediate, and this latency can be described by a response function
φ(t − ti ),
which on the basis of gure 1 we postulate
to be a long-memory process of the form with
0 < θ < 1.
φ(t) ∼ 1/t
1+θ
,
Using this, we can describe the rate of
views as a self-excited Hawkes conditional Poisson process that depends on all past events
λ(t) = V (t) +
X
µi
µi φ(t − ti )
(1)
is the number of potential viewers in uenced by
When the network is not ripe , corresponding to the case
hµi i
lowing a burst of activity depend on the susceptibility of the underlying social network to a particular video. We can therefore use the dynamic signature as a way of distinguishing on the basis of the exponent of the
is less than 1, then the activity generated by an
exogenous event does not cascade beyond the rst few generations, and the activity is given by
1 Abare (t) ∼ (t − tc )1+θ
treme cases: viral videos, quality videos, and junk. In this context, viral videos are those with precursory word-of-mouth growth resulting from epidemic like prop-
a viewer at time ti and V (t) captures all spontaneous views that are not triggered by network effects. when
count dynamics implies that the relaxation signatures fol-
power law governing their relaxation between three ex-
i|ti <t where
Classi cation of Content As outlined above, the existence of memory in the view
agation through a social network, characterized by an exponent (1 − 2θ ). Quality videos are similar to viral videos, but experience a sudden burst of activity rather than a bottom-up growth, and because of the quality of their content, subsequently trigger an epidemic cascade through the social network, relaxing with an exponent (1−θ ). Lastly, junk videos are those that experience a burst of activity for some reason (spam, chance, etc) but do not spread through the social network. Therefore their activity is determined largely by the
(2)
rst-generation of viewers, and they should relax as (1 + θ ).
If instead the network is ripe for a particular video, then the bare response is renormalized as the spreading is propagated through many generations, and the theory predicts the activity to be described as
Aexo (t) ∼
1 (t − tc )1−θ
(3)
If in addition to being ripe , the burst of activity is not the result of an exogenous event, but is instead fueled by endogenous growth, the bare response is renormalized in a different way giving
Aendo (t) ∼
1 (t − tc )1−2θ
(4)
While these results strictly hold for an ensemble of timeseries because of the stochasticity involved, we nd a surprisingly large number of individual videos that seem to obey these power-law relaxations exactly. Examples of this are shown in gure 2, suggesting that we can apply this formalism to individual videos.
Figure 3: Exponents for videos grouped by the fraction of views contained in their most active day (peak) relative to the total.
This is a natural way of separating endogenous
from exogenous videos, since the former have signi cant precursory growth, thus lowering the fractional weight contained in the peak.
Figure 3 shows the distribution of exponents obtained by grouping videos based on the fraction of views contained in their peak relative to the total. Videos experiencing an exogenous shock should have a very high percentage because there is little precursory growth, which is opposite for the endogenous case. Immediately one sees that based on this very simple criterion, the videos naturally fall into separate exponent classes, and we can extract
θ = 0.4 based on this
picture. A nal interesting result is that this classi cation does not rely on the magnitude of the largest peak, implying that identi cation of content can be made for large communities as well as more specialized, niche communities.
References Sornette, D.; DeschÂ&#x2C6; atres, F.; Gilbert, T.; and Ageon, Y. 2004.
Endogenous versus exogenous shocks in complex
networks: an empirical test using book sale ranking. Phys. Rev. Lett. 93(22):228701.